Enhancing Trading Strategies With Order Book Signals PDF

Enhancing Trading Strategies with Order Book SignalsI
Alvaro Carteaa , Ryan Donnellyb , Sebastian Jaimungalc
a
University of Oxford, Oxford, United Kingdom
b
Ecole Polytechnique Federale de Lausanne (EPFL), Lausanne, Switzerland
c
University of Toronto, Toronto, Canada
Abstract
We use high-frequency data from the Nasdaq exchange to build a measure of volume
imbalance in the limit order book (LOB). We show that our measure is a good predictor
of the sign of the next market order (MO), i.e. buy or sell, and also helps to predict price
changes immediately after the arrival of an MO. Based on these empirical findings, we
introduce and calibrate a Markov chain modulated pure jump model of price, spread, LO
and MO arrivals, and volume imbalance. As an application of the model, we pose and
solve a stochastic control problem for an agent who maximizes terminal wealth, subject
to inventory penalties, by executing trades using LOs. We use in-sample-data (January
to June 2014) to calibrate the model to ten equities traded in the Nasdaq exchange, and
use out-of-sample data (July to December 2014) to test the performance of the strategy.
We show that introducing our volume imbalance measure into the optimization problem
considerably boosts the profits of the strategy. Profits increase because employing our
imbalance measure reduces adverse selection costs and positions LOs in the book to take
advantage of favorable price movements.
Keywords: order imbalance, algorithmic trading, high-frequency trading, order flow,
market making, adverse selection
I
AC acknowledges the research support of the Oxford-Man Institute for Quantitative Finance and
the hospitality of the Finance Group at Sad Business School. SJ would like to thank NSERC and
GRI for partially funding this work. The authors would like to thank seminar participants at SIAM
Financial Mathematics and Engineering (Chicago), TMU Finance Workshop, Institut Louis Bachelier,
AMaMeF and Swissquote Conference, Imperial College Finance and Stochastics Seminar, and Morgan
Stanley (London) for comments on an earlier version of this article.
Email addresses: alvaro.cartea@maths.ox.ac.uk (Alvaro Cartea), ryan.donnelly@epfl.ch
(Ryan Donnelly), sebastian.jaimungal@utoronto.ca (Sebastian Jaimungal)
Electronic copy available at: http://ssrn.com/abstract=2668277

1. Introduction
The rise in computer power and the dominance of electronic exchanges have paved the
way to a surge in computerized trading algorithms. These algorithms are programmed to
trade in and out of positions and to handle different types of exposures such as inventory
risk and adverse selection costs. The development of many trading algorithms starts
with a preconceived strategy which is framed as an optimization problem, and then its
implementation is handled by computers. This approach has given way to a vast range of
algorithms. These are tailored to different objectives and employed in different trading
environments and markets, but common to all is that their performance depends on one
of the most basic building blocks in any trading strategy: how are orders scheduled, i.e.
how much and when to trade, and what order type should be employed.
How an algorithm determines the timing and type of order to execute a trade depends on
what the model predicts about price movements and what type of orders other agents
are sending to the market. In order driven markets, the main order types are limit
orders (LOs) and market orders (MOs). Agents LOs show an intention to buy or sell
an amount of the asset at a displayed price, and these rest in the limit order book (LOB)
until they are filled by an incoming MO or are cancelled by the agent who posted it.
MOs, on the other hand, are sent to the market and immediately executed against the
LOs resting in the book.
In this paper we use Nasdaq data to show how to employ LOB information in trading
algorithms. We use the volume posted on both sides of the LOB for a number of stocks
to build a measure of volume imbalance which proxies buying and selling pressures in the
market. We show that our measure of imbalance acts as a strong predictor of the rate
of incoming MOs as well as the direction and magnitude of price movements following
an MO. Given the arrival of an MO when volume imbalance is buy-heavy (sell-heavy),
there is a high probability that this MO is a buy (sell) order. Furthermore, immediately
following a buy (sell) MO, the magnitude and sign of the midprice is large and positive
(negative) when volume imbalance is buy-heavy (sell-heavy).
The ability to form accurate predictions of trade types and price changes is valuable
information that an agent can use to optimize her trading strategy. In general, incorpo-
rating volume imbalance in algorithmic trading models will improve the performance of
strategies. Using signals from the LOB helps to execute directional trades using MOs,
and tilt the resting orders in the LOB, to take advantage of favorable price movements
and to reduce adverse selection costs.
Furthermore, we use Nasdaq data to analyze the size of MOs relative to the volume
posted at the best bid and ask, and to analyze the financial performance of market
making strategies that only post at the best bid and ask. First, we find that depending
Electronic copy available at: http://ssrn.com/abstract=2668277

on the specific equity, between 91.6% and 99.9% of all MOs are filled with the volume
posted only at the best price. This means that LOs posted in the book deeper than
the best price have less than 0.1% to 8.4% chance of being executed by any given MO.
Second, we show that market making strategies that attempt at earning the spread by
always placing two-sided LOs at the best quotes suffer greatly from adverse selection to
the point where compensation of the spread does not allow positive expected profits.
These two findings shed light into the potential performance of the market making mod-
els that decide how deep LOs are posted in the LOB. For large tick stocks, those which
regularly trade with a bid-ask spread of one tick, most market making models proposed
in the literature result in strategies that mimic a market maker who is always posting
at-the-touch, this includes strategies that control exposure to inventory risk, see e.g.
Avellaneda and Stoikov (2008), Gueant et al. (2012), Fodra and Labadie (2012), Cartea
and Jaimungal (2015), and Cartea et al. (2014). Strategies that optimize expected
profits by choosing optimal depths require a level of granularity which is often too fine
relative to the tick size of the stock. In this case it is clear that to implement the market
making strategy the optimal depths must be rounded to the nearest tick which means
that postings are at the best bid or best ask, or at least one tick into the book. However,
postings deeper than the best quote are hardly ever filled so these strategies perform
very close to one that is always posting at the best quotes, and as we show they are
very likely to generate losses.
Our work is related to the microstructure and algorithmic trading literature that studies
the information conveyed by the LOB and order flow (i.e. MO activity). The early
empirical study of Biais et al. (1995) analyzes the dynamics of the LOB and MOs using
French data, and the theoretical work of Foucault (1999) presents a model of price
formation and order placement decisions in the LOB. More recently, with access to
better quality data, Cont et al. (2013) employ trade and quote data from Nasdaq to
perform a statistical study of the price impact of order book events. They show that
price changes are driven by order flow imbalances. The impact of volume imbalance on
price changes and trade arrivals is also studied in Lipton et al. (2013), and they develop
a model for the joint dynamics of the length of the best bid and ask queue and the arrival
of MOs. In Huang et al. (2015), the full LOB is modelled as a Markov queuing system
where interactions on the dynamics at several price levels are possible. They propose
their model as a market simulation tool to test the performance of trading algorithms.
In the context of algorithmic trading, Stoikov and Waeber (2012) consider an asset
liquidation problem where they employ LOB information to construct a measure of the
instantaneous supply and demand imbalance in the market. More recently, Bechler and
Ludkovski (2014) employ order flow information to develop optimal execution models
that take into account market impact and informational costs. Finally, Cartea and
Jaimungal (2014a) and Cartea and Jaimungal (2014b) employ ultra-high frequency data
3
to show that there is a positive relation between net order flow, defined as the difference
between the volume of buy and sell MOs, and prices of stocks and show how to develop
trading algorithms to target VWAP and execute a large number of shares.
The rest of this paper is structured as follows. In Section 2 we introduce our measure of
volume imbalance. We investigate the performance of a simple market making strategy
using data from Nasdaq, the results of which provide strong motivation for modifying
the strategy based on the volume imbalance process. In Section 3 we introduce a dy-
namic LOB model which reflects the relationship between the volume imbalance process
quantities such as arrival and size of MOs, and midprice dynamics. We also propose
and solve a trading problem using techniques of stochastic optimal control and provide
a verification proof. In Section 4 we examine the performance of the investment strategy
on the Nasdaq Exchange. Section 5 concludes, and the appendices contain proofs and
parameter estimates used in the paper.
2. Volume Imbalance: Order Arrival and Price Revisions
In this section we introduce and discuss aspects of our measure of volume imbalance.
This measure is simple to calculate and easy to incorporate in any algorithmic trading
strategy. In this paper we focus on equities that are traded in order driven markets (so
the LOB is visible), but in principle, measures based on quantities that show buying
and selling pressure can be incorporated into algorithms in a similar way to the trading
algorithm that we frame as a stochastic control problem in Section 3.
We define volume imbalance at time t as
Vtb Vta
t = [1, 1] , (1)
Vtb + Vta
where Vtb and Vta are the volumes at time t of LOs posted at the best bid and best ask
respectively. Clearly, when t is close to 1 there is strong buying pressure and when it
is close to 1 there is strong selling pressure.
Volume imbalance is a key quantity because it summarizes agents willingness to buy or

sell assets. This information can be used to predict the arrival and direction of MOs,
and help to predict the sign and magnitude of price changes we return to this point in
Subsection 2.2. Before discussing in more detail the attributes and statistical properties
of this quantity, we first illustrate how an investment strategy performs in the absence
of any measure that accounts for buying and selling pressure.
4
2.1. An example of trading strategy
A traditional investment strategy is to post on both sides of the LOB with the objective
of making roundtrip trades and earning the quoted difference between the LOs. A simple
version of this strategy is that of a market maker who is always posted at the best bid
and best ask. In this section, we test this strategy for two Nasdaq equities during
the period July to December 2014: Intel Corporation (INTC) and Oracle Corporation
(ORCL). These stocks exhibit behavior in which the spread, i.e. difference between best
offer and best bid, is equal to one cent for a significant portion of the trading day. In
Table 2 we report the results for nine other Nasdaq stocks.
We begin by studying the performance of a zero-intelligence strategy that posts LOs at

the best bid and ask in an attempt to earn the spread on roundtrip trades. We backtest
this strategy on historical data under two scenarios which differ in the volume of the
agents LOs.
Scenario 1. The agent imposes an inventory constraint such that inventory q

always satisfies |q| Q. If q 6= Q (q 6= Q), every market buy (sell) order changes
the agents inventory position by 1 (+1), and her wealth increases (decreases)
by an amount equal to the best ask (bid) price. At the end of the trading period,
any remaining inventory of the agent is liquidated using and MO which is filled at
the reigning best bid or ask price.
Scenario 2. As above, the agent imposes the constraint |q| Q. When a market
buy (sell) order arrives, the agents inventory position decreases (increases) by
min{V, Q + q} (min{V, Q q}), where V is the volume of the MO, and her wealth
increases (decreases) by the traded volume multiplied by the best ask (bid) price.
At the end of the trading period, any inventory of the agent is liquidated at the
reigning best bid or ask price via an MO.
The difference between the scenarios is that in Scenario 1 the LOs are for one unit of
the asset and in Scenario 2 the LOs are for a number of shares large enough to meet
the full volume of the incoming MO. Moreover, in both scenarios we assume that each
trading period is 30 minutes long, but we exclude the first and last trading period in
each day. We conduct these tests on each trading day from June to December, 2014, so
there are 125 days, each with 11 trading periods during the day.
The left panel of Figure 1 shows the annualized mean and standard deviation of the zero-
intelligence strategy under Scenario 1. The leftmost point on each curve corresponds to
Q = 1 with unit increments to Q = 10. Increments then increase by 10 up to Q = 200.
The right panel shows the Sharpe ratio for a range of values of the inventory constraint.
5
1000 10
INTC INTC
500 ORCL ORCL
0
0
Sharpe Ratio
Expextation
500 10
1000 20
1500
30
2000
2500 40
0 100 200 300 400 0 50 100 150 200
Standard Deviation Maximum Inventory
Figure 1: Performance of best bid/ask LO strategy under Scenario 1. Each point represents a different
value of Q. The leftmost point on each curve corresponds to Q = 1 with unit increments to Q = 10.
Increments then increase by 10 up to Q = 200.
The results for Scenario 2 are shown in Figure 2. Here the values of Q are scaled by
a factor of 100 from those of Scenario 1 so the total range is 100 to 20000. The much
larger magnitude of earnings under Scenario 2 are due to larger individual trade sizes.
Clearly, the expected earnings of the zero-intelligence strategy are negative under both
scenarios. Any agent who provides liquidity to the market in this naive way will be
driven out of business due to the strategys exposure to severe adverse selection costs.
An attempt to alleviate adverse selection costs is to post LOs deeper in the LOB, hoping
that MOs walk beyond the best quote to fill these LOs. However, this adjustment will
have a negligible effect on the expected profits because LOs resting at worse prices than
the best quotes are unlikely to be filled. Consider the size of an MO relative to the
volume at the best bid or offer. In Table 1, we show the number of market buy and
sell orders that only touch the best quotes in the LOB, and the number that touch LOs
beyond the best quote. It is clear that a disproportionate number of MOs only involve
LOs at the best quotes for these two stocks, approximately 0.1% of the MOs are large
enough to walk beyond the best quote.
Although the strategy employed by the market maker is extremely simple, many of the
market making and investments strategies in the extant literature exhibit a very similar
behavior to the one discussed here for large tick stocks, see for instance Avellaneda and
Stoikov (2008), Gueant et al. (2012), Fodra and Labadie (2012), Cartea and Jaimungal
(2015), and Cartea et al. (2014) large tick stocks are those that generally trade with
a bid-ask spread of one tick. The trading strategies developed in these papers use a
continuous control variable for the price at which the trader or market maker posts
LOs. These optimal depths are the output of a stochastic optimal control problem
6
5
x 10
1 10
INTC INTC
0.5 ORCL ORCL
0
0
Sharpe Ratio
Expextation
0.5 10
1 20
1.5
30
2
2.5 40
0 1 2 3 4 0 50 100 150 200
Standard Deviation x 10
4 Maximum Inventory
Figure 2: Performance of best bid/ask LO strategy under Scenario 2. Each point represents a different
value of Q. The leftmost point on each curve corresponds to Q = 100 with increments of 100 up to
Q = 1000. unit increments to Q = 10. Increments then increase by 1000 up to Q = 20000.
Best Quote Only Beyond Best Quote P(VM O VLO )

Buys Sells Buys Sells
INTC 35,595 38,451 54 50 0.999
ORCL 30,001 27,502 41 45 0.999
Table 1: Number of MOs that touch or go beyond the best quote. Data are taken from a full month
of trading (January, 2014). The column labelled P(VM O VLO ) is the probability that an MO has
smaller volume than all limit orders posted at the best price, and hence only engages the best quote.
where the main driver to adjust the depth is inventory position and end of trading
horizon. When the strategies are implemented, the optimal depths must be rounded to
the nearest tick size, but with the market behavior demonstrated in Table 1 a continuous
control price seems unreasonable (see also Table B.7 in the appendix for similar data
on other stocks). The effect of this mismatch between a continuous control, and the
grid on which asset prices move in exchanges, is more pronounced for large tick stocks
because the probability of an MO walking beyond the best quote is essentially zero.
When the continuous controls are rounded to an appropriate multiple of the tick size,
either all of the values are rounded to the best quote, or they are rounded to a tick which
has very little activity in the market. Thus, if the posting strategy based on continuous
controls is inventory based, then the resulting strategy essentially becomes one of always
posting at the best bid and ask, with an inventory constraint as the scenarios tested
above. On the other hand, models that are not purely inventory based, for instance
Cartea et al. (2014), the strategy is modified based on the observation of recent order
flow, so rounding the optimal strategy to the nearest tick results in a strategy that
posts sometimes at the touch, but also sometimes deeper in the book, and imposes an
inventory constraint.
7
4 4
x 10 x 10
3 2.5
Market Buys Market Buys
2.5 Market Sells Market Sells
Number of Trades
Number of Trades
2
2
1.5
1.5
1
1
0.5
0.5
0 0
SellHeavy Neutral BuyHeavy SellHeavy Neutral BuyHeavy
Imbalance Level Imbalance Level
(a) INTC (b) ORCL
Figure 3: Number of MOs that are buys and sells depending on imbalance level. Imbalance levels
correspond to subintervals [1, 0.33), [0.33, 0.33], and (0.33, 1]. Data are taken from a full month
of trading in January 2014.
2.2. Volume imbalance in order driven equity markets
The zero-intelligence strategy described above, as well as those that behave in an es-
sentially identical manner, are a few examples of algorithms whose performance can be
enhanced if volume imbalance is employed. In general, if a strategy employs additional
information which predicts trade arrivals and price movements then it is able to protect
itself from adverse selection costs and take advantage of price movements.
To show the predictive power of our volume imbalance measure we first divide the
imbalance measure interval [1, 1] into three subintervals, referred to as buy-heavy
when t 31 , 1 , sell-heavy when t 1, 31 , and neutral when t 31 , 13 . Figure

3 shows that depending on which subinterval imbalance t lies in, the type of incoming
MO, buy or sell, can be established with high accuracy.
Furthermore, the empirical distribution of the midprice change following an MO varies

depending on which subinterval was occupied at the time of the MO. In panels (a) and
(b) of Figure 4 we show this empirical distribution conditioned on imbalance regime. In
this figure, a time lag of 10 ms is used between the arrival of the MO and the calculation
of the price change. The positive (negative) bias of price changes after an MO in a buy-
heavy (sell-heavy) regime is not simply due to the fact that it is much more likely to
observe a market buy (sell) order in that regime. If we compute the price change when
only observing market buy orders, we obtain the distribution shown in panels (c) and
(d) of Figure 4. These two figures lend support to the idea that the direction of the price
change after a market buy order is most of the time positive (see for example Cartea
and Jaimungal (2014b)), and the magnitude of the change has a clear dependence on
8
4
x 10
2
Sell-Heavy 12000 Sell-Heavy
Neutral Neutral
Buy-Heavy 10000 Buy-Heavy
1.5
8000
1
6000
4000
0.5
2000
0 0
3 2 1 0 1 2 3 3 2 1 0 1 2 3
Midprice Change (ticks) Midprice Change (ticks)
(a) INTC all MOs (b) ORCL all MOs
12000 10000
Sell-Heavy Sell-Heavy
10000 Neutral Neutral
Buy-Heavy 8000 Buy-Heavy
8000
6000
6000
4000
4000
2000
2000
0 0
3 2 1 0 1 2 3 3 2 1 0 1 2 3
Midprice Change (ticks) Midprice Change (ticks)
(c) INTC buy MOs (d) ORCL buy MOs
Figure 4: Distribution of midprice change 10 ms after an MO. Subintervals and data are equivalent to
those in Figure 3.
the level of volume imbalance.
We choose a 10 ms window between MOs and price changes because LO activity increases
immediately following an MO to such a large magnitude that this increase in activity can
be thought of as having been caused by the MO the results in the paper are not altered
if the window is chosen to be up to 100 ms. We note that our statistical results do not
provide conclusive evidence on the causality between MO and LO activity. However,
the framework we develop here does not require a deeper understanding of causality, it
suffices to observe that LO activity is greatest right after the arrival of MOs.
To highlight this increased LO activity post MO activity, panels (a) and (b) of Figure
5 shows the proportion of three types of LO events that occur after MO arrivals. The
horizontal axis is the time window immediately after the MO event and the vertical axis
9
is proportion of all LO events accounted for in this time window. The three types of
LO events counted in the figure are:
Type 1: placed or cancelled.
Type 2: placed or cancelled at a price equal to or better than the current best bid
or ask.
Type 3: placed or cancelled and causes the midprice to change.
The difference between a Type 2 order and a Type 3 order is that a Type 2 order does
not necessarily have to change the midprice of the asset. If a limit buy order is placed
at the best bid, or if an existing order a the best bid is cancelled (but other orders still
remain at that price), then this would be a Type 2 order but not a Type 3 order. The
inclusions Type 3 Type 2 Type 1 hold.
Panel (c) in Figure 5 shows the proportion of the trading day accounted for immediately
following MO arrivals. The calculation of these proportions are performed as follows:
let [0, T ] be the time interval of interest, panel (c) is all of January 2014 with the first
and last 30 minutes of each day removed (we assume it is one connected interval by
concatenating trading hours together so that T = 415800 seconds). Let {tk }N k=1 be the
i Ni
sequence of times when MOs occur, and let {k }k=1 be the sequence of times when LO
events of type i occur. Let t > 0 represent the lag time after an MO for which its effect
on LO activity may be significant. For each value of the lag, t, define the union of time
intervals It = N k=1 [tk , tk +t). The quantity It represents the amount of time in [0, T ]
i i
that takes place less than t after an MO. Finally, define Pt = card({ki }N
k=1 It ).
This quantity represents the number of LO events of type i which occur within t after
i
an MO. Panels (a) and (b) plot Pt /N i as a function of t for each type of LO event,
and panel (c) plots m(It )/T as a function of t for the two equities of interest, where
m is the Lebesgue measure.
As Figure 5 shows, a disproportionate amount of all LO activity occurs shortly following

an MO. To put in perspective the significance of the disproportion, note that less than
0.5% of the trading day is accounted for in the short time (50 ms) following all MOs.
However, between approximately 40% and 90% of all LO activity (whether defined by
type 1, 2 or 3) is accounted for in the same time period.
The combination of using a time lag of 10 ms in computing midprice changes and the
behavior shown in Figures 4 produces a notable feature in the distribution of price
changes. Most of the mass in these distributions occurs at the 0 and 1 tick levels,
with less probability of observing a 0.5 tick movement. Consider the scenario when
the spread is one tick and an incoming MO fills the full volume at the best quote. Soon
10
100 100 0.5
INTC
Proportion of Events (%)
Proportion of Events (%)
Proportion of Time (%)

ORCL
80 80 0.4
60 60 0.3
40 40 0.2
20 type 1 20 type 1 0.1

type 2 type 2
type 3 type 3
0 0 0
0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50
Lag (ms) Lag (ms) Lag (ms)
(a) INTC (b) ORCL (c)
Figure 5: Proportion of all LO events that occur shortly after an MO (panels a and b). Proportion of
trading day as percentage accounted for immediately following MO (panel c). Data are taken from a
full month of trading in January 2014.
after, there are three possible events: i) no new LOs are added within the new spread,
ii) an LO is added within the new spread to replenish the order that was filled, or iii)
an LO is added within the new spread, but on the other side of the book. If the original
MO is a buy (sell), the midprice change in each scenario is i) 0.5 (0.5) ticks, ii) 0
ticks, or iii) 1 (1) tick, respectively. The drastic increase in LO activity immediately
following an MO along with the fact that INTC and ORCL are large tick stocks explains
why scenario i) is much more rare than scenarios ii) or iii).
The above empirical results clearly show that volume imbalance conveys important
information about arrival rate of MOs and the distribution of prices. Here we have
examined a simple way to summarize LOB information on buying and selling pressure,
but depending on the particular asset, or how agents post intentions to sell or buy
using passive orders in other types of markets, it might be useful to synthesize volume
imbalance information in a different way.
Moreover, regardless of the way in which buying and selling pressure is synthesized,
trading algorithms could be considerably improved if this information is included. For
example, in market making algorithms, or those where the investor aims at profiting
from roundtrip trades, it is clear that basing her trading on volume imbalance protects
the strategy from adverse selection costs and increases expected profits as a result of
better inventory management and by tilting the LOs to take advantage of expected
moves in prices.
11
3. Trading Algorithm with Volume Imbalance Information
In this section we develop a trading algorithm where the objective is to maximize profits
by posting LOs to make roundtrip trades. While the LOB dynamics we propose are
appropriate over a wide range of equities, we tailor the control processes to large tick
stocks and assume that LOs are posted only at the best bid and/or the best offer for
small tick stocks the strategy needs to be adjusted so that it is possible to post LOs
within the spread. In Subsection 3.1 we present model ingredients such as the dynamics
for: MOs, midprice, and volume imbalance. In Subsection 3.2 we present and solve
the investors investment problem. In Subsection 3.4 we discuss the optimal trading
strategy. Finally, we devote Section 4 to discussing the performance of the trading
algorithm using market data for INTC and ORCL.
3.1. Market orders, midprice dynamics, volume imbalance
We work with a complete filtered probability space (, F, P). Let E = R+ R3 , E be the

Borel sets of E, and l , + , and be three doubly stochastic Poisson random measures
(PRMs) on (E, E) with compensators l , + , and . These PRMs count events in
the book due to LO activity (corresponding to l ) and MO activity (corresponding to
), and thus drive all of the dynamics in the model. More details on their structure
are outlined in subsection 3.1.1. Below we describe how the components of the LOB,
including LOs, MOs, midprice, and spread evolve.
Market Orders: The number of market buy (sell) orders that have occurred up to
time t is denoted by Mt+ (Mt ). These quantities are
Z tZ
Mt = (dy, du) .
0 yR3
Limit Order Events: We are only concerned with an LO placement or cancellation

if it causes the midprice, spread, or imbalance regime (defined below) to change. If we
refer to these as Type 4 events, then with the description of other types as in Section
2.2 we have Type 3 Type 4 Type 2. We denote the number of Type 4 events as
Z tZ
Mtl = l (dy, du).
0 yR3
12
Midprice: The midprice is modelled by the pure jump process
Z tZ
St = S0 + y1 (l + + )(dy, du) , (2)
0 yR3
where the value of y1 R represents the changes in midprice due to market activity.
Volume Imbalance: Rather than modelling the imbalance process t directly, which
can be rather noisy, we instead consider a finite state imbalance regime process Zt
{1, . . . , nZ }. This process acts as an approximation, (and smoothing) to the true value
of imbalance by dividing the interval [1, 1] into nZ subintervals (labelled such that
subinterval 1 corresponds to the largest negative values (i.e. the sell heavy regime), and
n corresponds to the largest positive values (i.e. the buy heavy regime) and Zt = k
corresponds to t lying within subinterval k. The process Zt is given by
Z tZ
Zt = Z0 + (y2 Zu )(l + + + )(dy, du) , (3)
0 yR3
where the value of y2 R represents the state of the imbalance regime process immedi-
ately following any market activity.
Spread: The spread between the best bid and best offer is modelled as a finite state
process which takes values in {1, . . . , n } and is given by
Z tZ
t = 0 + (y3 u )(l + + + )(dy, du) , (4)
0 yR3
where the value of y3 R represents the value of the spread immediately following any
market activity.
3.1.1. Properties of the doubly stochastic PRMs
The compensators of l , + , and are denoted by l , + , and and are chosen to

be of the form i (dy, dt) = i (Zt , t )FZi t ,t (dy)dt, where for any (Z, )
Z
i
FZ, (dy) = 1 . (5)
yR3
Furthermore, l , + , and are independent conditional on (Zt , t ), specifically
1 j
p
i
lim P(Mt+h Mti = 1 Mt+h Mtj = 1 | Zt , t ) = i,j i (Zt , t )j (Zt , t ) , (6)
h0 h
13
where i,j is the Kronecker delta and where i and j are any of l, +, or . This property
ensures that any single change in St , Zt , and t is a result of only one PRM. If the distri-
i
butions FZ, are chosen to have support only on y2 {1, . . . , nZ } and y3 {1, . . . , n },
then from the definitions in (3) and (4) we have Zt {1, . . . , nZ } and t {1, . . . , n }
with probability 1. Furthermore, (6), along with the factored form of the compensators,
makes the vector process (Zt , t ) a continuous-time Markov chain. In the sequel it may
occasionally be convenient to refer to the pair imbalance and spread as a single state in
which case it is represented by J = (Z, ).
Proposition 1 (Generator of (Zt , t )). The generator matrix G of the process J t =

(Zt , t ) is given by
+
GJ,K = l (J ) TJ,K
l
+ + (J ) TJ,K + (J ) TJ,K , for J 6= K ,
(7)
+
(J ) 1 TJ,J

GJ,J = l (J ) 1 TJ,J
l
+ (J ) 1 TJ,J ,
for J {1, . . . , nZ } {1, . . . , n }, where

Z
i
TJ,K = 1(y2 ,y3 )=K FJi (dy) .
yR3
Proof See Appendix A.1.
l
Finally, the distribution FZ, (dy) is chosen to have support only on the two hyperplanes
y3
y1 = 2 . This is because if a single LO placement or cancellation causes the spread to
change, then the midprice must also necessarily change by an amount which is exactly

half of the change in the spread. We do not impose this constraint on FZ, , even
though the change in spread and midprice due to a single MO event must have the same
relationship. The reason for not imposing this constraint is that immediately following
an MO, there is an increase in the intensity of LO activity which has an effect on the
spread and midprice. This extra LO activity (as depicted above in Figure 5) can be
viewed as a result of the MO immediately prior, and so the changes in spread and
midprice can be attributed to the MO arrival itself.
3.2. Agents optimization problem
As discussed above, see for example Table 1 (and Table B.7 in Appendix), we observe
that MOs hardly ever walk beyond the best quote this is a common feature in large
14
tick stocks. Thus, here we assume that the agent posts LOs only at-the-touch, i.e. the
best bid or best ask. In reality, the agent has LOs posted at many prices in the LOB
deeper than the best bid/ask so that if those prices were to become the best bid or ask
in the future, the agents orders at that time are closer to the front of the queue. This
is also why we assume that the agents LOs posted at the best bid or ask are always
executed by an incoming MO. It is possible to generalize the results to account for a
probability of being filled which is state dependent and less than 1, but the qualitative
behavior of the optimal strategy does not change.
The agent begins with a certain amount of wealth x0 which changes as MOs arrive and
fill her LOs and her cash process X = (Xt )0tT satisfies the SDE

t t
+
dXt = t St + +
dMt t St dMt , X0 = x 0 ,
2 2
where the processes = (t )0tT , which take on values 0 or 1, are the agents
controls for whether she is posting buy and/or sell orders. Recall that Mt are the
counting processes for market buy and sell orders.
As well, the agents inventory q = (qt )0tT satisfies the SDE
dqt = t+ dMt+ + t dMt .
We consider an agent who aims to find the strategy which maximizes expected
terminal wealth subject to additional penalties. In particular, her value function is
given by
Z T
2

H(t, x, q, S, J ) = sup E XT + qT (ST `(qT , T )) qu du Ft , (8)
(s )tsT A t
where T is the terminal time of the strategy, `(qT , T ) is a liquidation penalty, 0 is

a running inventory penalty parameter, and A is the set of admissible strategies which
are Ft -predictable such that the inventory remains bounded between finite upper and
lower bounds: Q qt Q. If qt = Q the agent does not post LOs on the sell side of
the book, i.e. must choose t+ = 0, and similarly, when qt = Q she chooses t = 0.
The liquidation penalty `(q, ): is increasing in q, its absolute value is increasing in

, and `(0, ) = 0. This penalty reflects the costs from liquidating terminal inventory
with a single MO (which may walk the LOB) and encourages the agent to implement a
strategy where qT is close to 0.
RT
The term t
qu2 du represents a running inventory penalty which discourages the agent
15
from holding non-zero inventory positions for significant lengths of time. The parameter
can be used as a risk control by determining the strength of this penalty. A large
value of makes the agent act in a more conservative manner by inducing the strategy
to unwind positions, long or short, very quickly. See Guilbaud and Pham (2013) and
Cartea and Jaimungal (2015) for similar inventory control terms, and Cartea et al.
(2014) who show that this running penalty arises in the context of ambiguity aversion
when the agent is not confident about the drift of the midprice.
3.3. Feedback control of the optimal strategy
To find the optimal strategy, consider the Hamilton-Jacobi-Bellman (HJB) equation

associated with the value function H of the control problem (8)
t Hq 2 +l (J ) E[Dl H | J ]+ sup + (J ) E[D++ H | J ]+ sup (J ) E[D H | J ] = 0 ,

+ {0,1} {0,1}
(9)
subject to the terminal condition H(T, x, q, S, J ) = x + q (S `(q, )). Here the oper-
ator Dl acts as follows:
Z
l
D H(t, x, q, S, J ) = H (t, x, q, S + y1 , y2 , y3 ) H (t, x, q, S, J ) l (dy, ) ,
yR3
and the expectation operator conditional on J acts as

Z
l
E[D H | J ] = H (t, x, q, S + y1 , y2 , y3 ) H (t, x, q, S, J ) FJl (dy) . (10)
yR3
Similarly, the operators D are defined by

Z

D H = H t, x
S
, q , S y1 , y2 , y3 H (dy, ) ,
yR3 2
where we have suppressed explicit dependence on (t, x, q, S, J ) for easier readability.

These have conditional expectation
Z

E[D H | J ] = H t, x
S , q , S y1 , y2 , y3 H FJ (dy) .

yR 3 2
(11)
Equation (10) represents the expected change in the value function due to the arrival of
an LO event (a new LO or cancellation of an existing order). Similarly, equations (11)
represent the expected change in the value function due to the arrival of a buy or sell
MO, depending on the agents strategy at the time of the arrival of the MO.
16
We make the ansatz H(t, x, q, S, J ) = x + q S + h(t, q, J ) where each term carries the
following interpretation: the cash position x, the book value of the shares which are
marked-to-market at the midprice, and h(t, q, J ) which represents the extra value of
optimally trading until the terminal date. Inserting this into the HJB (9), after a
number of tedious computations, leads to a system of equations for h:
t h q 2 + l (J ) ql (J ) + l (t, q, J )

+ + + + +
+ sup (J ) + (q ) (J ) + + (t, q, J )
+ {0,1} 2
(12)

+ sup (J ) (q + ) (J ) + (t, q, J ) = 0 ,
{0,1} 2
h(T, q, J ) = q `(q, ) ,
where
X
i (J ) = y1 FJi (y1 , y2 , y3 ) ,
y1 ,y2 ,y3
X
l (t, q, J ) = l
(h(t, q, K) h(t, q, J )) TJ,K ,
K
X

(t, q, J ) =
(h(t, q , K) h(t, q, J )) TJ,K .
K
The terms i (t, q, J ) represent the expected change in the value function due to state
transitions. The terms i (J ) represent the expected midprice change after the arrival
of an event. This identification, allows us to interpret the remaining contributions to
equation (12).
Terms of the form q i (J ) represent the instantaneous expected change of the value of
the agents holdings due to market activity of type i (where i can be l, +, or ). If
this product is positive, then the agent is in a state where her inventory holdings are
expected to increase in value. If it is negative, then the current market state dictates her
inventory holdings are expected to decrease in value. The terms (J ) represent
the immediate costs incurred due to adverse selection after having an LO filled (note
that (J ) is generally positive according to Figure 4 and the signs of the PRMs in (2)).
Finally, the terms 2 are the immediate profits made by the agent due to the arrival
of an MO. Consequently, the relative sizes of 2 and (J ) are an important indicator
of the agents strategy. We discuss this point below when analyzing the optimal trading
strategy.
Proposition 2 (Existence of solution). Equation (12) together with its terminal con-
dition has a unique classical solution.
17
Recall that the control processes are restricted to values of 0 or 1. From (12), the
resulting optimal feedback controls are given by a simple comparison:

1, 2 + (J ) + + +
1 (t, q, J ) > 0 (t, q, J ) , and q 6= Q ,
+
(t, q, J ) = (13a)
0, otherwise ,

1,
2
(J ) +
1 (t, q, J ) > 0 (t, q, J ) , and q 6= Q ,

(t, q, J ) = (13b)
0, otherwise .

Due to the discrete nature of the controls, there is little hope of finding a closed-form
solution for the value function, however, (12) can be easily solved numerically.
The form of the optimal controls in (13a) and (13b), together with some reasonable
foresight about the qualitative behavior of the optimal strategy, reveals a simple rule
of thumb. It is expected that for a fixed state of imbalance and spread, J , there
is a boundary in the space of the variables (t, q) such that if q lies above (below) this
boundary, then the agent optimally posts a limit sell (buy) order. The sign of 2 (J )
plays a large part in determining whether this boundary lies above or below the curve
q = 0. The terms
1 (t, q, J ) and 0 (t, q, J ) act as corrections based on the future
liquidation penalty and running inventory penalty. If there is no inventory constraint or
penalization, and the time to maturity is very large, the terms
1 (t, q, J ) and 0 (t, q, J )
are equal and the strategy is determined by the sign of 2 (J ). This last expression
is the expected amount of cash that the agent earns if the position is unwound at the
midprice.
Theorem 3 (Verification Theorem). Let h be the solution to (12) and define

H(t, x, q, S, J ) = x + q S + h(t, q, J ). Then the candidate solution H equals the value
function H as defined in (8).
3.4. Optimal Posting Strategy
To show a typical posting strategy, we select a model with 3 states in the volume
imbalance regime process, and the stock trades with a spread of 1 or 2 ticks, i.e. nZ = 3
18
and n = 2. For a full set of parameters, see the Appendix C.1. These parameters
result from the estimation procedure outlined in Section 4.1 when applied to 30 minutes
of INTC data starting at 10:00 on January 24, 2014. Parameters of most interest are
shown below. The components of these vectors represent states, J = (Z, ) with the
following convention: the first three components correspond to = 1 and Z = 1, 2, 3.
The last three components correspond to = 2 and Z = 1, 2, 3.
+ = ( 0.038 0.121 0.600 1.711 1.188 3.042 ),

= ( 0.600 0.121 0.038 3.042 1.188 1.711 ),
+ = ( 0.240 0.306 0.713 0.111 0.152 0.469 ),
= ( 0.713 0.306 0.240 0.469 0.152 0.111 ).
These parameters indicate the intensity of MOs and expected midprice change imme-
diately following an MO for each state. Intensities are displayed in units of second1
and midprice changes are in ticks. Note that higher imbalance regimes (i.e. the book is
tilted to the buy side) generally indicate a higher frequency of market buy orders and
also larger expected (positive) midprice changes after buy MOs. High imbalance also
indicates lower frequency of market sell orders and smaller jump sizes after sell orders.
This is consistent with market behavior shown in Figures 3 and 4.
Figure 6 shows the optimal strategy corresponding to sell and buy LOs. In the top panel
the spread is = 1 tick and in the bottom panel is = 2 ticks. In both panels, the
light grey region corresponds to posting a limit sell order, the dark grey region a limit
buy order, and the black region indicates an overlap where both a buy and sell order are
posted. Finally, the white region indicates that no order should be placed in the book.
There are several features in this figure to note. As imbalance becomes more buy-
heavy, the buy (sell) region becomes larger (smaller) in anticipation of the asset price to
increase. Limit sell orders become less desirable because they are more likely to incur
immediate losses due to adverse selection. Any desire to sell the asset in a buy-heavy
imbalance regime is due to the liquidation penalty, to be paid at time T , and the running
penalty, which add up to a cost that outweighs the gains from trading in the direction
of the expected increase in prices.
In both the buy-heavy and sell-heavy regimes, the boundary between the sell and no
sell regions generally lies considerably above or below, respectively, the zero inventory
level. Note that in the sell-heavy regime the boundary is significantly below the level
of zero inventory. This is consistent with rule of thumb, discussed after Proposition 2,
to sell the asset in the sell-heavy regime when 2 + (J ) > 0 because the earnings of
the half-spread are larger than the expected price increase immediately following the
MO. Similarly, in the buy-heavy regime, the boundary is significantly above the zero
19
50 50 50
Place Buy Place Buy Place Buy
Inventory Level Place Sell Place Sell Place Sell
Inventory Level
Inventory Level
0 0 0
50 50 50
0 200 400 600 0 200 400 600 0 200 400 600
Time Time Time
Sell-Heavy Neutral Buy-Heavy
= 1 tick
50 50 50
Place Buy Place Buy Place Buy
Place Sell Place Sell Place Sell
Inventory Level
Inventory Level
Inventory Level
0 0 0
50 50 50
0 200 400 600 0 200 400 600 0 200 400 600
Time Time Time
Sell-Heavy Neutral Buy-Heavy
= 2 ticks
Figure 6: Dark grey (light grey) region is where buy (sell) LOs are posted. Black region is overlap
where both buy and sell orders are be posted. White region indicates no-trade region, where the agent
does not have LOs in the book.
inventory level. In this case it is usually not optimal to post a sell order because the
gains from having the limit sell order filled are lower than the expected increase in
midprice, 2 + (J ) < 0. In this region the agent is better off not posting because on
average she loses the half-spread she earns to adverse selection costs.
For a fixed imbalance regime, the sell region is larger for larger values of . This implies
that within a fixed imbalance regime, the agent is more willing to post orders when the
spread is large. This does not always have to be the case, but for the particular set of
parameters chosen note that 2 + (J ) is larger for = 2 ticks than for = 1 tick.
Finally, the top panel of the figure shows two large no-trade regions in the sell-heavy and
buy-heavy imbalance regimes. In these regions, the agent withdraws from the market.
It is optimal to let current inventory appreciate in value because the expected future
value of what the agent is holding outweighs the earnings from an additional trade net
of any inventory costs (running penalty or terminal liquidation). The agent remains
withdrawn from the market until the imbalance or spread change, or until the passage
of time takes the strategy close enough to the terminal date T , so the future cost of
having to cross the spread to liquidate outweighs the benefit of inventory appreciation.
20
We also observe a small no-trade region in the neutral regime when inventory is close
to zero and the strategy approaches the terminal date T .
4. Performance of Strategy on the Nasdaq Exchange
In this section we analyze the performance of the strategy on trade data from the Nasdaq
exchange for INTC and ORCL. Our data set consists of all messages sent to Nasdaqs
LOB during all trading days in 2014. We employ the first six months of data to calibrate
model parameters and test the strategys performance over the following six months.
We assume that the agents trading window is half-hour intervals. That is, the agent
sets a horizon of 30 mins over which she trades and closes all positions at the end of
the window. We record the PnL of the strategy for every half-hour and compute the
Sharpe ratio for a range of values of the running inventory parameter .
The choice of a 30 minute trading horizon is arbitrary and may have an effect on the
profitability of the strategy. Some stocks are more active than others over that window so
the chances of making roundtrip trades varies, thus one could choose the trading window
based on the arrival rate of MOs and the average of traded volume. Alternatively, it is
possible to sow together the various 30 minute periods to create a strategy that trades all
day long and only unwinds at the end of the day. Such a strategy should outperform the
one presented here. Nonetheless, the approach taken here allows us to demonstrate the
historical value of adding imbalance as a state variable, and also allows us to generate
many return periods from which we can compute meaningful estimates of returns and
variability of returns.
We organize the rest of this section as follows. In Subsection 4.1 we show how the model
parameters are estimated. In Subsection 4.2 we discuss how to forecast and update the
model parameters which are used in the execution of the out-of-sample strategy. Finally,
in Subsection 4.3 we present the out-of-sample performance of the strategy.
4.1. Estimating Model Parameters
To compute the optimal trading strategy, we solve (12) and then use (13a) and (13b)
to find in which regions LOs are posted. To solve these equations we do not require
all model parameters. Rather, for each Z and and each i = +, , l, we only require
the quantities i (J ), i (J ) = y1 ,y2 ,y3 y1 FJi (y1 , y2 , y3 ), and TJ,K
i
= y1 FJi (y1 , K). The
P P
value of i (J ) is the rate at which an event of type i occurs when the current imbalance
and spread are equal to Z and respectively, recall that J = (Z, ). The value of i (J )
is the expected change in midprice immediately after an event of type i when the event
i
occurs in state J . Finally, the value of TJ,K is the probability of transitioning from
21
state J to state K after an event of type i. The sum over y1 represents averaging out
all possible midprice changes so the remaining quantity represents transitions between
imbalance and spread only.
Estimating the model parameters over a time window is a straightforward procedure.

Moreover, to ensure that the model does not contain any long term speculation based
on the average growth of the asset midprice, we impose symmetry constraints between
complementary states of imbalance. For example if rate of arrival of buy MOs is higher
than arrival of sell MOs then a long term strategy is to buy and hold the asset because
the model predicts that in the long term prices will increase.
The complementary imbalance state of Z is defined to be Z = nZ Z + 1. We also use

the short-hand notation J = (Z, ). The constraints we impose are the following:
l (J ) = l (J ) , (14a)

(J ) = (J ) , (14b)
l (J ) = l (J ) , (14c)
(J ) = (J ) , (14d)
l l
TJ,K = TJ, K
, (14e)

TJ,K = TJ,K . (14f)
We remark that we impose this symmetry to preclude long term speculation strategies
and acknowledge that many high-frequency trading strategies are designed to benefit
from short-lived trends in the midprice. There is no contradiction in a model where
the long term growth of the midprice is symmetric as discussed here, but also exhibits
short term deviations around the midprices long term trend. For example there are
models that seek alpha, see Cartea et al. (2014), and models that take advantage of
how innovations in order flow affect the trend in prices, see for instance Cartea and
Jaimungal (2014b), Cartea and Jaimungal (2014a).
Let NJi denote the number of events of type i that occurred from 0 to T when the
imbalance and spread are equal to J . Also let J be the total occupation time of state
J so that J J = T . Then i (J ) is estimated by
P
l
NJl + NJl
NJ + NJ
(J ) = , and (J ) = .
J + J J + J
iNJi
Further, let yJi1 , . . . , yJ be all the midprice changes that occur following an event of
22
type i from state J . Then
PNJl PNJl PNJ k
PNJ
lk
k=1 yJ
lk
k=1 yJ k=1 yJ + k=1 yJk
l (J ) = , and (J ) = .
NJl + NJl NJ + NJ
i
Finally, let NJ,K be the number of times that the state transitions from J to K. Then
l l
NJ,K + NJ, K
NJ,K + NJ, K
l
TJ,K = , and TJ,K = .
NJl + NJl NJ + NJ
Note that these quantities satisfy (14a) to (14f). These are also the expressions that
result from a maximum likelihood estimation with the constraints imposed. The pa-
rameters used in the numerical example of the previous section were obtained through
this procedure, and so by construction they satisfy the constraints which results in the
symmetry between buy and sell regions seen in Figure 6.
4.2. Forecasting Model Parameters
Testing the trading strategy on out-of-sample data requires forecasts of model parame-
ters for each half-hour interval based on the estimated parameters from previous intervals
recall that the trading horizon is T = 30 mins. We begin by labelling each interval
according to the day of the year and the time of day in which it takes place. The data
consist of all the messages sent to the Nasdaq exchange in all trading days in 2014,
giving a total of 249 days consisting of 13 half-hour intervals, i.e. T = 30 mins, in each
trading day. We identify each interval as the pair (m, n) where m denotes the day in the
year and n denotes the half-hour of the day, thus m {1, . . . , 249} and n {1, . . . , 13}.
We employ the first 6 months of data, i.e. the first 124 days, to calibrate all model
parameters for each half hour in the day. These parameter estimates are used to develop
a predictive model so that model parameters in a given 30 minute period (over which
we aim to backtest out-of-sample) are estimated from the data we already observed in
the last 30 minute period. We do this in two stages described in detail below.
Let im,n = (im,n (1), . . . , im,n (nZ, ))0 and im,n = (im,n (1), . . . , im,n (nZ, ))0 , where nZ, =
nZ n , and 0 is the transpose operation. In the first stage, for each trading period in
the first six months of the year, January to June (m = 1, . . . , 124), we estimate all of
i
the model parameters m,n and im,n . Thus, we have in-sample estimates of the model
parameters (for the first 6 months), and a different model for each day and each trading
period.
In Figures 7 and 8 we show motivation for the forecasting method we propose below.
23
0.8 0.8 0.8
0.7 0.7 0.7

Market Order Intensity

0.6 0.6 0.6
0.5 0.5 0.5
0.4 0.4 0.4

0.3 0.3 0.3
0.2 0.2 0.2
0.1 0.1 0.1
0 0 0
9AM 10AM 12PM 1PM 2PM 9AM 10AM 12PM 1PM 2PM 9AM 10AM 12PM 1PM 2PM
Time Time Time
(a) Sell-Heavy (b) Neutral (c) Buy-Heavy
Figure 7: Intraday seasonality of MO intensity for ORCL. Each curve is a different day in the month
of January 2014. Each point on a single curve is the intensity of market buy orders conditional on the
state of imbalance over a 30 minute period. The thick black curve is the average intensity within the
same window of all days in the sample. Similar behavior is shown for INTC.
When calibrating the model parameters over each 30 minute period, we observe a well
known intraday seasonality effect. Figure 7 shows the arrival intensity of MOs, condi-
tional on the imbalance regime, which is generally higher at the beginning and end of
the day giving a characteristic U shape. Moreover, Figure 8 shows the size of midprice
changes following an MO. The pattern is generally high at the beginning of the day and
low at the end of the day giving a characteristic S shape.
In the second stage, we assume a factor model that relates the parameters of one trading
interval n to the parameters of the previous trading interval n 1 on the same day m:
i i i i i
m,n = n + n m,n1 + m,n , (15)
i i i
im,n = n + n im,n1 + m,n , (16)
i i i i
where n , n RnZ, , n , n RnZ, nZ, , and all of the idiosyncratic error terms
i i
m,n and m,n are independent with mean zero.
Using the model parameters estimated in-sample for the first 6 months, we then perform
i i i
a multilinear regression to obtain the estimates n and n , and the factor loadings n
i
and n .
To test the strategys performance, the next step is to forecast the parameters that
are used in the out-of-sample period. For day m we employ the data in the first 30
minute period (i.e. n = 1 which is the first half hour of the day) to estimate the model
i
parameters m,1 and im,1 , but do not trade over this first half-hour of the day. We then
use the regression model to predict the model parameters for the next period (in which
i
we trade). Denote these forecasts by m,2 and im,2 , which are the parameters used to
24
1 1 1
0.8 0.8 0.8
0.6 0.6 0.6

Jump Size
Jump Size
Jump Size
0.4 0.4 0.4
0.2 0.2 0.2
0 0 0
0.2 0.2 0.2
0.4 0.4 0.4

9AM 10AM 12PM 1PM 2PM 9AM 10AM 12PM 1PM 2PM 9AM 10AM 12PM 1PM 2PM
Time Time Time
(a) Sell-Heavy (b) Neutral (c) Buy-Heavy
Figure 8: Intraday seasonality of midprice changes following an MO for ORCL. Each curve is a different
day in the month of January 2014. Each point on a single curve is the average midprice change following
a market buy order (measured in ticks) conditional on the state of imbalance over a 30 minute period.
The thick black curve is the average midprice change within the same window of all days in the sample.
Similar behavior is shown for INTC.
formulate the trading strategy over the second 30 minute period. After the second 30
minute period has passed, we use the observed data for the second period to estimate
i
m,2 and im,2 , and once again use the regression model to forecast the parameters of the
third period (over which we trade), and so on. Succinctly, we use the estimates resulting
from the regression (15) and (16) to forecast the parameters for half-hour n = 2, . . . , 131
in day m using the recursion:2
i i i i
m,n = n + n m,n1 ,
i i
im,n = n + n im,n1 .
We do not perform this type of regression and forecasting for the transition probabilities
i
TJ,K,m,n because numerical investigation has shown that the trading strategy depends
much more heavily on i and i . Instead, the value of TJ,K,m,n
i
is set equal to the mean
i
of TJ,K,m,n taken over all m {1, . . . , 124}.
1
Although we are able to use our forecasting method to obtain parameters for n = 13, we do not
test the trading strategy in the last 30 minute interval as market behaviour is notably different towards
the end of the day.
2
After computing the forecast of im,n it is possible we obtain a negative value. In this case, we set
m,n = 102 . We do not set it equal to zero because we do not want to impose that a state is absorbing.
i
We also impose a maximum value on each im,n of 10. This is solely for purposes of numerical stability
when solving equation (12) and does not have a significant qualitative impact on the resulting optimal
strategy.
25
4.3. Out-of-sample Performance of Trading Strategy
Now that we have a recursive method to forecast model parameters for the interval
(m, n), based on the parameters of interval (m, n 1), we test the out-of-sample per-
formance of the trading strategy for all days in July to December 2014. Our forecast
method does not allow us to test the strategy in the first half-hour of each day, and
we also exclude testing in the last half-hour of each day as market behavior is notably
different towards the end of the trading day. This leaves us with a total of 1375 half-hour
intervals to perform an out-of-sample test of the strategy. At the end of each half-hour
trading period, any non-zero inventory is liquidated by crossing the spread, consistent
with the use of the penalty term `(q, ) = sign(q) 2 , which assumes that there is enough
liquidity at the best quote to fill the agents MO.
Here we fix the maximum inventory constraint at Q = Q = 50 and choose a range

of values of the running penalty parameter to see how it affects the PnL. Moreover,
we examine how the strategy performs when volume imbalance is modelled assuming
nZ = 1, 3, 5 states, and fixing n = 2. Recall that volume imbalance t [1, 1] and
that this is divided into nZ subintervals, and n is the grid on which the spread lives.
When nZ = 1 there is only one volume imbalance state and the midprice can either be
in a state where the spread is one or two ticks.
Figure 9 shows the annualized mean PnL versus annualized standard deviation of the
resulting wealth over all trading intervals for INTC (left panel) and ORCL (right panel).
As expected, the worst performance is when nZ = 1 and the results for nZ equal to 3
and 5 are approximately the same. For both stocks we observe that increasing the value
of the running penalty parameter increases the performance of the strategy: expected
PnL increases and its standard deviation decreases.
Figure 10 shows another perspective of the results. It depicts the Sharpe ratio of the
strategy as a function of the parameter , and the risk-free rate is zero. We observe that
as the agent enforces stricter controls on inventory by increasing , the Sharpe ratio of
the strategy also increases.
The results in Figures 9 and 10 should be compared to those of the zero-intelligence

strategy. It is clear that employing the information impounded in the volume imbalance
process, i.e. MO arrival and price innovations, considerably boosts the profits of an
investment strategy based on roundtrip trades. The sources of the profits stem from
protecting the strategy from adverse selection costs, and positioning the LOs to take
advantage of price movements. In addition, we also observe that imposing a running
inventory penalty enhances the profitability of the strategy as in Guilbaud and Pham
(2013) and Cartea and Jaimungal (2015).
26
500 500
nZ = 1 nZ = 1
nZ = 3 nZ = 3
nZ = 5 nZ = 5
0 0
Mean
Mean
500 500
1000 1000
0 20 40 60 80 100 0 10 20 30 40 50
Standard Deviation Standard Deviation
(a) INTC (b) ORCL
Figure 9: Annualized Expectation vs. Standard Deviation of trading strategies based on different
number of imbalance states and Q = Q = 50. Each point represents a different value of , ranging
from 0 to 105 (larger values of correspond to smaller values of standard deviation).
50 50
nZ = 1 nZ = 1
40 nZ = 3 40 nZ = 3
nZ = 5 nZ = 5
30 30
Sharpe Ratio
Sharpe Ratio
20 20
10 10
0 0
10 10
20 20
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
x 10
5 x 10
5
(a) INTC (b) ORCL
Figure 10: Annualized Sharpe ratio vs. of trading strategies based on different number of imbalance
states and Q = Q = 50, and the risk-free rate is zero.
27
In Table 2 we show the annualized Sharpe ratio for the zero-intelligence strategy when
applied to a set of large tick stocks for various values of the inventory constraint Q
(these results take place under Scenario 1 as discussed in Section 2.1). These values
show similar behaviour to what is seen in Figure 1 indicating the failure to make positive
profits through roundtrip trades when always posting LOs on both sides of the LOB.
In contrast, Tables 3, 4, and 5 show the annualized Sharpe ratio over the same set of
equities when employing our strategy with volume imbalance and spread information.
The results are for nZ = 1, 3, 5, and the spread is 1 or 2 ticks, i.e. n = 2. We observe a
large boost in performance across the board. The results in Table 3 do not incorporate
observations of imbalance in the strategy since nZ = 1, but executions are dependent on
the spread. This explains the relatively poor performance when compared to the results
in Tables 4 and 5. Further, the Sharpe ratio generally increases with an increase in the
running inventory parameter .
Maximum Inventory Q
1 2 4 10 20 40 100 200
AA -29.62 -33.77 -31.57 -21.89 -15.00 -10.18 - 6.97 - 6.41
AMAT -34.20 -37.90 -33.66 -22.27 -15.15 -11.60 - 8.81 - 7.47
ARCC -14.46 -16.07 -14.11 -11.28 - 9.13 - 6.95 - 5.30 - 4.85
BXS -33.87 -27.24 -22.25 -17.15 -14.49 -12.12 -11.96 -11.96
CSCO - 8.50 -18.03 -20.56 -15.71 -11.24 - 9.26 - 7.04 - 6.18
EBAY -17.11 -20.55 -21.21 -18.08 -13.36 - 8.58 - 5.55 - 4.17
FMER -33.12 -34.11 -28.45 -18.63 -13.44 -11.30 -10.25 -10.22
IMGN -49.92 -40.68 -29.62 -18.39 -13.70 -10.54 - 8.33 - 7.90
INTC -20.97 -26.15 -26.53 -23.11 -17.37 -11.98 - 8.24 - 6.90
NTAP -45.96 -44.68 -37.70 -23.71 -15.08 - 9.75 - 8.06 - 8.04
ORCL -20.24 -28.23 -28.36 -20.89 -13.59 - 9.17 - 6.71 - 5.53
Table 2: Annualized Sharpe ratio of zero-intelligence trading strategies based on various values of Q.
So far the strategy assumes that the agents LOs are always at the front of the queue.
This assumption is unrealistic and introduces a bias in the profitability of the trading
strategy. We address this issue by assigning a fill probability to the LOs. This probability
depends on the type of MO and the state of imbalance immediately prior to the MO.
Define the two vectors
p+ = ( 0.20 0.35 0.50 0.65 0.80 ) ,

p = ( 0.80 0.65 0.50 0.35 0.20 ) ,
which represent fill probabilities depending on the state of imbalance and the type of
MO. If 15 , 15 indicating a neutral market in which Z = 3, then the agents LO

is filled with probability p+ (3) = p (3) = 0.5. If 15 , 35 indicating a buy-heavy

market in which Z = 4, then a buy (sell) MO fills an agents posted sell (buy) LO with
28
Inventory Penalty Parameter
7
0 10 2 107 4 107 106 2 106 4 106 105
AA -3.78 - 6.42 - 5.93 - 5.06 - 3.68 - 0.71 3.00 8.75
AMAT -6.09 - 8.65 - 8.84 - 8.64 - 7.73 - 5.28 - 1.74 4.80
ARCC -4.94 - 8.41 - 8.52 - 9.06 - 8.73 - 7.16 - 4.09 - 0.53
BXS 1.50 2.68 3.29 3.78 4.49 5.21 5.81 6.01
CSCO -4.18 - 8.81 - 9.71 -10.23 -10.06 - 8.65 - 5.92 - 0.49
EBAY 0.90 1.27 1.76 2.50 3.86 5.24 7.07 10.28
FMER 1.51 6.40 7.99 9.82 12.90 16.26 20.08 23.81
IMGN 1.40 - 0.91 - 1.08 - 1.25 - 1.22 - 1.20 - 0.85 - 0.07
INTC -7.52 -16.04 -16.69 -17.19 -17.36 -16.24 -14.43 - 9.42
NTAP 0.50 1.82 1.81 1.51 1.88 2.60 3.17 4.99
ORCL 0.88 5.36 7.13 9.67 13.13 16.65 20.58 26.30
Table 3: Annualized Sharpe ratio of trading strategies based on various values of , the number of
imbalance states here is nZ = 1 and n = 2 and the maximum inventory constraint is Q = 50.

0 107 2 107 4 107 106 2 106 4 106 105
AA 6.11 9.24 11.13 12.74 17.82 22.10 24.90 28.32
AMAT 5.32 10.03 12.09 14.62 17.95 20.99 24.56 28.64
ARCC - 0.85 - 0.41 0.19 0.73 2.10 3.44 5.39 8.18
BXS 2.66 2.86 3.11 3.68 3.93 4.70 6.49 7.96
CSCO 8.19 10.11 11.67 13.86 17.99 21.04 24.82 30.78
EBAY 3.13 5.51 6.28 7.49 9.43 11.17 14.14 18.17
FMER 5.31 7.00 7.92 9.30 12.05 14.45 17.02 22.02
IMGN 2.08 1.65 1.62 1.67 1.98 1.92 2.20 3.60
INTC 1.90 8.04 9.94 12.51 15.96 18.70 21.38 25.13
NTAP 1.98 2.32 2.33 2.73 3.10 4.34 5.88 7.64
ORCL 11.65 16.76 18.74 20.86 23.71 26.94 29.17 34.25
probability p+ (4) = 0.65 (p (4) = 0.35). If 35 , 1 indicating a strongly buy-heavy

market in which Z = 5, then a buy (sell) MO fills an agents posted sell (buy) LO with
probability p+ (5) = 0.80 (p (5) = 0.20). The probabilities for imbalance states Z = 1
and Z = 2 are defined similarly.
Table 6 shows the Sharpe ratio of the strategy with nZ = 5 and when fill rates depend
on type of MO and the state of imbalance as described. As expected, the Sharpe ratios
of the strategy are lower than those shown in Table 5, where the only difference is
the fill rates, but the ratios are still considerable higher than those obtained with the
zero-intelligence strategy.
29
7
0 10 2 107 4 107 106 2 106 4 106 105
AA 6.12 8.35 9.79 12.06 17.34 20.33 23.99 27.93
AMAT 6.81 9.49 10.91 12.79 16.53 19.67 23.72 28.49
ARCC -0.66 -0.01 0.73 1.40 3.07 4.35 5.85 9.34
BXS 4.25 4.50 4.78 5.08 6.45 6.43 7.57 9.42
CSCO 9.48 11.60 12.91 15.38 18.94 22.73 27.14 32.85
EBAY 3.41 4.59 6.00 7.74 9.96 12.42 16.12 20.91
FMER 5.24 6.57 7.02 8.05 10.48 12.69 15.77 20.71
IMGN 2.83 2.75 2.73 2.43 2.05 2.06 2.49 3.15
INTC 4.49 9.68 11.25 12.85 15.60 17.93 20.88 24.45
NTAP 1.63 2.29 2.63 2.73 2.74 3.79 4.53 7.32
ORCL 11.75 14.28 16.09 18.32 21.90 25.20 28.99 33.29

0 107 2 107 4 107 106 2 106 4 106 105
AA 2.42 3.96 4.92 4.46 6.83 10.85 12.81 17.28
AMAT 3.90 4.17 4.66 5.83 8.99 8.94 12.88 17.83
ARCC -2.22 -2.62 - 2.23 - 2.63 - 1.26 - 1.27 - 0.68 2.03
BXS 1.83 1.89 1.48 0.68 1.00 1.92 2.45 1.62
CSCO 6.46 7.68 7.56 10.00 11.95 14.12 17.18 20.79
EBAY 1.99 2.71 5.11 4.92 3.84 6.34 7.77 11.02
FMER 1.42 2.96 2.78 4.05 5.61 5.48 7.46 9.65
IMGN 0.79 0.47 0.19 0.11 1.54 0.28 1.52 2.08
INTC 0.43 5.23 5.87 7.63 9.96 9.66 12.57 15.91
NTAP -0.22 1.62 - 0.57 1.89 - 0.47 - 0.34 0.70 0.81
ORCL 7.87 9.68 10.65 11.71 13.00 15.87 15.78 21.14
Table 6: Annualized Sharpe ratio of trading strategies with modified fill probabilities depending on the
level of imbalance. The trading strategy here is based on a number of imbalance states equal to nZ = 5
and n = 2 and the maximum inventory constraint is Q = 50.
We finish this section by adding a few remarks about the strategys performance. First,
for the purposes of this historical test, and indeed for the entire optimization problem
as stated, here we focused on an agent who utilizes LOs on both sides of the market.
The performance of the strategy is expected to improve if the agent also uses MOs
to take advantage of price innovations which are anticipated by the volume imbalance
process. Moreover, the use of volume imbalance can be implemented in a vast range of
other optimal execution problems. For example, algorithms to acquire/liquidate a large
number of shares, as well as Pairs Trading algorithms.
30
Finally, we acknowledge that measures based on the volume of shares resting in the LOB,
or number of LOs in the LOB, may be subject to manipulation. In most equity markets
it is costless to cancel or amend an LO and although illegal, market participants may
artificially tilt the LOB by submitting LOs on one side of the book, only to withdraw
them quickly enough whilst taking advantage of traders that employ strategies like the
one described, and those suggested, in this paper. This illegal activity is commonly
referred to as spoofing and very difficult to detect in modern electronic exchanges
where a large number of market participants employ different trading strategies and
some can trade at extremely low latencies. In the particular cases discussed here, one
cannot tell if throughout 2014 the stocks we study are subject to spoofing at any point
in time. Even if this is the case, the statistical properties of volume imbalance as a
predictor of MO activity and price innovations seem robust to this manipulation of LOs
which would certainly erode the predictive power of volume imbalance.
5. Conclusions
We employ message data from the Nasdaq exchange to build a measure of volume
imbalance and show that it predicts type of MO arrival (buy or sell) as well as predict
price innovations. This measure captures buying and selling pressure in the LOB. In
particular we show that when the LOB is buy-heavy the probability of the next MO
being a buy order is much higher than it being a sell order. Moreover, we also show
that when the LOB is buy-heavy, price revisions after MO arrival are on average large
and positive, i.e. larger than the average price revision seen when the LOB is neither
buy-heavy nor sell-heavy. We observe the same empirical behavior when the book is
sell-heavy. This relationship between volume imbalance and market order activity is
consistent across a large sample of Nasdaq stocks we examined.
This measure is simple to build and can be employed by a wide range of trading algo-
rithms. A clear effect of using this information in a trading model is to reduce adverse
selection costs and to take advantage of favorable price movements.
As an example of an algorithmic trading strategy we solve an optimal investment prob-

lem of an agent who provides liquidity to the LOB. The agents objective is to maximize
expected terminal wealth by completing roundtrip trades whilst penalizing inventory
positions. We use trade data to show the out-of-sample performance of the strategy
during the period 1 July to 31 December 2014. We demonstrate that including the
volume imbalance process considerably boosts the strategys process and the Sharpe
ratio of the strategy is considerable higher than that obtained by most models proposed
in the literature, i.e. Avellaneda and Stoikov (2008), Gueant et al. (2012), Fodra and
Labadie (2012), Cartea and Jaimungal (2015), and Cartea et al. (2014).
31
Appendix A. Proofs
Appendix A.1. Proof of Proposition 1
Proof of Proposition 1 Define the process Ft = E[(J T )|Ft ] = f (t, J t ) for some
function to be specified later. Using Itos Lemma and the fact that Ft is a martingale,
we have that f satisfies the following equation
Z X
i i
t f (t, J ) + f (t, y2 , y3 ) f (t, J ) (J )FJ (dy) = 0
yR3 i
XZ X
i i
t f (t, J ) + f (t, y2 , y3 ) f (t, J ) 1(y2 ,y3 )=K (J )FJ (dy) = 0
K yR3 i
X Z X
i
t f (t, J ) + f (t, K) f (t, J ) 1(y2 ,y3 )=K (J )FJi (dy) =0
yR3
K
X i X
i i
t f (t, J ) + f (t, K) f (t, J ) (J )TJ,K = 0.
K i
Letting [f (t)]J = f (t, J ) and using the matrix G introduced in the proposition, the
matrix form of this equation is
t f + Gf = 0 ,
which has solution
f (t) = eG(T t) .
If (J ) is defined as the indicator function for a particular state J 0 , then f (t) is a vector
of transition probabilities to state J 0 from each arbitrary state J over a time of T t.
This is the same expression which dictates transition probabilities for a continuous time
Markov chain with generator matrix G.
Appendix A.2. Proof of Proposition 2
Proof of Proposition 2 Equation (12) can be written in the form:
t h = F (h) ,
where F is a vector valued function of a vector argument. Each component of F is a

piecewise linear function of h, the coefficients of which depend on the regions defined
by the inequalities in (13a) and (13b). Since inventory, imbalance, and spread are all
bounded, the number of regions defined by the inequalities in (13a) and (13b) is finite.
Therefore, F is Lipschitz, and equation (12) has a unique classical solution.
32
Appendix A.3. Proof of Theorem 3
Proof of Theorem 3 Let h be the solution to equation (12) and define a candidate
optimal value function H(t, x, q, S, J ) = x + qS + h(t, q, J ). From Itos Lemma we have
Z T

H(T, XT , qT , ST , J T ) = H(t, x, q, S, J ) + t h(u, Xu , qu , Su , J u )du
t
Z T Z

+ qu y1 + h(t, qu , y2 , y3 ) h(t, qu , J u )l (dy, du)
t yR3
Z T Z
u
+ u+ + (qu u+ )y1 + h(t, qu u+ , y2 , y3 ) h(t, qu , J u )+ (dy, du)
t yR3 2
Z T Z
u
+ u (qu + u )y1 + h(t, qu + u , y2 , y3 ) h(t, qu , J u ) (dy, du) .
t yR3 2
Taking an expectation conditional on Ft on both sides and rearranging yields

Z T

H(t, x, q, S, J ) = E H(T, XT , qT , ST , J T )Ft E
t h(u, qu , J u )du
t
Z TZ

+ qu y1 + h(t, qu , y2 , y3 ) h(t, qu , J u ) l (dy, du)
t yR3
Z T Z
u
+ u+ + (qu u+ )y1 + h(t, qu u+ , y2 , y3 ) h(t, qu , J u ) + (dy, du)
t yR3 2
T
Z Z
u

+ u (qu + u )y1 + h(t, qu + u , y2 , y3 ) h(t, qu , J u ) (dy, du)Ft .
t yR3 2
(A.1)
Equation (12) then yields the inequality

Z T
2

H(t, x, q, S, J ) E H(T, XT , qT , ST , J T ) (qu ) duFt
t
Z T

2

= E H(T, XT , qT , ST , J T ) (qu ) duFt
t
Z T
2

= E XT + qT (ST `(qT , T )) (qu ) duFt .
t
Since this inequality holds for arbitrary controls , we have:

Z T
2

H(t, x, q, S, J ) sup E XT + qT (ST `(qT , T )) (qu ) duFt
(s )tsT A t
= H(t, x, q, S, J ) . (A.2)
33
Now, if is selected to be of the form in (13a) and (13b), then (A.1) implies:
Z T
2

H(t, x, q, S, J ) = E H(T, XT , qT , ST , J T ) (qu ) duFt
t
Z T
2

= E H(T, XT , qT , ST , J T ) (qu ) duFt
t
Z T
2

= E XT + qT (ST `(qT , T )) (qu ) duFt
t
Z T
2

sup E XT + qT (ST `(qT , T )) (qu ) duFt
(s )tsT A t
= H(t, x, q, S, J ) . (A.3)
Combining (A.2) and (A.3) yields the result.
Appendix B. Market Behavior Across Equities
In this section we show that several equities demonstrate the same behavior which is
illustrated in Table 1 and Figures 3 and 4. We begin with Table B.7 which shows
how many market orders are filled only at the best posted price versus how many walk
through more than one full level of the book. We see that GOOG has the smallest
percentage or orders remaining strictly within the best price at 91.6%, but this number
is often higher than 99%. Generally, stocks that can be considered large tick exhibit a
larger percentage of trades strictly within the best price.
Table B.8 shows the average trade intensities for buy and sell market orders depending
on the state of volume imbalance. Generally we see that the frequency of buy (sell)
orders increases as imbalance becomes more buy-heavy (sell-heavy).
Finally, in Table B.9 we show the average midprice change 10 ms after an MO that
occurred within each state of imbalance. Once again we see a clear tendency for the
midprice change to have a larger (smaller) magnitude after a market buy (sell) order
when imbalance is more buy-heavy.
34
First Tick Only Beyond First Tick P(VM O VLO )
Buys Sells Buys Sells
AAPL 100,362 105,655 4,581 4,527 0.958
FARO 1,745 2,374 64 109 0.960
GOOG 32,096 34,969 3,085 3,075 0.916
INTC 35,595 38,451 54 50 0.999
MMM 22,996 25,745 130 118 0.995
NTAP 28,519 27,118 104 123 0.996
ORCL 30,001 27,502 41 45 0.999
SMH 3,087 3,084 7 4 0.998
Table B.7: Number of MOs that touch only the first tick or go beyond the first tick. Data is taken from
a full month of trading in January, 2014 (first and last 30 minutes of each day removed). The column
labelled P(VM O VLO ) is the probability that an MO has smaller volume than all limit orders posted
at the best price, and hence only engages the best quote.
Average Buy Intensity Average Sell Intensity

Sell-Heavy Neutral Buy-Heavy Sell-Heavy Neutral Buy-Heavy
AAPL 0.223 0.257 0.276 0.293 0.271 0.228
FARO 0.005 0.005 0.003 0.008 0.007 0.004
GOOG 0.060 0.090 0.112 0.102 0.095 0.071
INTC 0.024 0.048 0.232 0.288 0.056 0.021
MMM 0.034 0.058 0.080 0.072 0.063 0.046
NTAP 0.031 0.050 0.167 0.133 0.045 0.031
ORCL 0.022 0.041 0.181 0.213 0.037 0.015
SMH 0.002 0.004 0.020 0.017 0.004 0.002
Table B.8: Average trade intensities within each state of imbalance. Data is taken from a full month
of trading in January, 2014 (first and last 30 minutes of each day removed).
Appendix C. Parameter Sets
The entire calibrated parameter set is available upon request.
Appendix C.1. Parameters of Section 3.4
The full set of parameters used to compute the optimal trading strategy in Figures 6
are given here. In this example, the number of imbalance regimes is nZ = 3 and the
number of spread values is n = 3 giving a total of nZ n = 6 states. These states are
ordered in the same convention as used previously: the first three states correspond to
= 1 and Z = 1, 2, 3 and the last three states correspond to = 2 and Z = 1, 2, 3.
35
Market Buy Order Market Sell Order
Sell-Heavy Neutral Buy-Heavy Sell-Heavy Neutral Buy-Heavy
AAPL 2.212 2.527 3.098 3.245 2.751 2.388
FARO 1.846 3.047 4.251 4.414 3.781 2.602
GOOG 8.115 8.866 10.619 10.678 9.191 8.432
INTC 0.044 0.211 0.457 0.494 0.208 0.057
MMM 1.134 1.507 1.613 1.716 1.496 1.177
NTAP 0.270 0.498 0.723 0.668 0.477 0.291
ORCL 0.031 0.290 0.570 0.603 0.376 0.128
SMH 0.262 0.411 0.759 0.730 0.450 0.217
Table B.9: Average midprice change (in units of ticks) after an MO within each state of imbalance.
Data is taken from a full month of trading in January, 2014 (first and last 30 minutes of each day
removed).
Due to the form of equation (12), the full form of each compensator is not required.
i
Rather, we only require i (J ), i (J ) = y1 ,K y1 FJi (y1 , K), and TJ,K = y1 FJi (y1 , K).
P P
l = ( 1.257 0.538 1.257 85.187 23.132 85.187 ),

+ = ( 0.038 0.121 0.600 1.711 1.188 3.042 ),
= ( 0.600 0.121 0.038 3.042 1.188 1.711 ),
l = ( 0.137 0.000 0.137 0.160 0.000 0.160 ),
+ = ( 0.240 0.306 0.713 0.111 0.152 0.469 ),
= ( 0.713 0.306 0.240 0.469 0.152 0.111 ).
36

0.00 0.73 0.00 0.05 0.21 0.01

0.50 0.00 0.50 0.00 0.00 0.00

0.00 0.73 0.00 0.01 0.21 0.05

Tl = ,

0.04 0.01 0.33 0.00 0.61 0.01

0.06 0.00 0.06 0.44 0.00 0.44
0.33 0.01 0.04 0.01 0.61 0.00

0.40 0.36 0.24 0.00 0.00 0.00

0.01 0.61 0.34 0.00 0.02 0.01

0.02 0.40 0.53 0.00 0.03 0.03

T+ = ,

0.00 0.33 0.22 0.33 0.11 0.00

0.02 0.26 0.04 0.00 0.67 0.00
0.00 0.69 0.25 0.00 0.00 0.06

0.53 0.40 0.02 0.03 0.03 0.00

0.34 0.61 0.01 0.01 0.02 0.00

0.24 0.36 0.40 0.00 0.00 0.00

T = ,.

0.25 0.69 0.00 0.06 0.00 0.00

0.04 0.26 0.02 0.00 0.67 0.00
0.22 0.33 0.00 0.00 0.11 0.33
References
Avellaneda, M. and S. Stoikov (2008, November). High-frequency trading in a limit

order book. Quantitative Finance 8, 217224.
Bechler, K. and M. Ludkovski (2014). Optimal execution with dynamic order flow
imbalance. arXiv preprint arXiv:1409.2618.
Biais, B., P. Hillion, and C. Spatt (1995). An empirical analysis of the limit order book
and the order flow in the paris bourse. The Journal of Finance 50(5), 16551689.
Cartea, A., R. Donnelly, and S. Jaimungal (2014). Algorithmic trading with model
uncertainty. SSRN eLibrary, http://ssrn.com/abstract=2310645.
Cartea, A. and S. Jaimungal (2014a). A closed-form execution strategy to target VWAP.

Available at SSRN 2542314.
Cartea, A. and S. Jaimungal (2014b). Incorporating order-flow into optimal execution.

Available at SSRN 2557457.
37
Cartea, A. and S. Jaimungal (2015). Risk metrics and fine tuning of high frequency
trading strategies. Mathematical Finance 25(3), 576611.
Cartea, A., S. Jaimungal, and J. Ricci (2014). Buy low sell high: A high frequency
trading perspective. SIAM Journal on Financial Mathematics 5(1), 415444.
Cont, R., A. Kukanov, and S. Stoikov (2013). The price impact of order book events.
Journal of Financial Econometrics.
Fodra, P. and M. Labadie (2012). High-frequency market-making with inventory con-

straints and directional bets. Working papers, HAL.
Foucault, T. (1999). Order flow composition and trading costs in a dynamic limit order
market. Journal of Financial markets 2(2), 99134.
Gueant, O., C.-A. Lehalle, and J. Fernandez-Tapia (2012). Dealing with the inventory
risk: a solution to the market making problem. Mathematics and Financial Economics,
131.
Guilbaud, F. and H. Pham (2013). Optimal high frequency trading with limit and
market orders. Quantitative Finance 13(1), 7994.
Huang, W., C.-A. Lehalle, and M. Rosenbaum (2015). Simulating and analyzing or-
der book data: The queue-reactive model. Journal of the American Statistical
Association 110(509), 107122.
Lipton, A., U. Pesavento, and M. G. Sotiropoulos (2013). Trade arrival dynamics and
quote imbalance in a limit order book. arXiv:1312.0514.
Stoikov, S. and R. Waeber (2012). Optimal asset liquidation using limit order book
information. Available at SSRN 2113827.
38

Enhancing Trading Strategies With Order Book Signals PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Enhancing Trading Strategies With Order Book Signals PDF

Uploaded by

Copyright:

Available Formats

Enhancing Trading Strategies with Order Book SignalsI

Alvaro Carteaa , Ryan Donnellyb , Sebastian Jaimungalc

Electronic copy available at: http://ssrn.com/abstract=2668277

Electronic copy available at: http://ssrn.com/abstract=2668277

2. Volume Imbalance: Order Arrival and Price Revisions

We define volume imbalance at time t as

Volume imbalance is a key quantity because it summarizes agents willingness to buy or

We begin by studying the performance of a zero-intelligence strategy that posts LOs at

Scenario 1. The agent imposes an inventory constraint such that inventory q

Best Quote Only Beyond Best Quote P(VM O VLO )

(a) INTC (b) ORCL

2.2. Volume imbalance in order driven equity markets

Furthermore, the empirical distribution of the midprice change following an MO varies

(a) INTC all MOs (b) ORCL all MOs

(c) INTC buy MOs (d) ORCL buy MOs

the level of volume imbalance.

Type 1: placed or cancelled.

Type 3: placed or cancelled and causes the midprice to change.

As Figure 5 shows, a disproportionate amount of all LO activity occurs shortly following

Proportion of Events (%)

Proportion of Time (%)

20 type 1 20 type 1 0.1

(a) INTC (b) ORCL (c)

3.1. Market orders, midprice dynamics, volume imbalance

We work with a complete filtered probability space (, F, P). Let E = R+ R3 , E be the

Limit Order Events: We are only concerned with an LO placement or cancellation

3.1.1. Properties of the doubly stochastic PRMs

The compensators of l , + , and are denoted by l , + , and and are chosen to

Furthermore, l , + , and are independent conditional on (Zt , t ), specifically

Proposition 1 (Generator of (Zt , t )). The generator matrix G of the process J t =

for J {1, . . . , nZ } {1, . . . , n }, where

Proof See Appendix A.1.

3.2. Agents optimization problem

As well, the agents inventory q = (qt )0tT satisfies the SDE

dqt = t+ dMt+ + t dMt .

where T is the terminal time of the strategy, `(qT , T ) is a liquidation penalty, 0 is

The liquidation penalty `(q, ): is increasing in q, its absolute value is increasing in

3.3. Feedback control of the optimal strategy

To find the optimal strategy, consider the Hamilton-Jacobi-Bellman (HJB) equation

t Hq 2 +l (J ) E[Dl H | J ]+ sup + (J ) E[D++ H | J ]+ sup (J ) E[D H | J ] = 0 ,

and the expectation operator conditional on J acts as

Similarly, the operators D are defined by

where we have suppressed explicit dependence on (t, x, q, S, J ) for easier readability.

Theorem 3 (Verification Theorem). Let h be the solution to (12) and define

Proof See Appendix A.3.

3.4. Optimal Posting Strategy

+ = ( 0.038 0.121 0.600 1.711 1.188 3.042 ),

4. Performance of Strategy on the Nasdaq Exchange

4.1. Estimating Model Parameters

Estimating the model parameters over a time window is a straightforward procedure.

The complementary imbalance state of Z is defined to be Z = nZ Z + 1. We also use

4.2. Forecasting Model Parameters

0.7 0.7 0.7

Market Order Intensity

Market Order Intensity

0.5 0.5 0.5

0.4 0.4 0.4

0.2 0.2 0.2

0.1 0.1 0.1

(a) Sell-Heavy (b) Neutral (c) Buy-Heavy

0.8 0.8 0.8

0.6 0.6 0.6

0.2 0.2 0.2